2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
37 #include <linux/module.h>
39 #include <linux/init.h>
40 #include <linux/slab.h>
41 #include <linux/kernel.h>
42 #include <linux/vmalloc.h>
44 #include <linux/if_arp.h> /* For ARPHRD_xxx */
49 #include <linux/jhash.h>
51 #include <net/addrconf.h>
52 #include <linux/inetdevice.h>
53 #include <rdma/ib_cache.h>
55 #define DRV_VERSION "1.0.0"
57 const char ipoib_driver_version[] = DRV_VERSION;
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("IP-over-InfiniBand net driver");
61 MODULE_LICENSE("Dual BSD/GPL");
63 int ipoib_sendq_size __read_mostly = IPOIB_TX_RING_SIZE;
64 int ipoib_recvq_size __read_mostly = IPOIB_RX_RING_SIZE;
66 module_param_named(send_queue_size, ipoib_sendq_size, int, 0444);
67 MODULE_PARM_DESC(send_queue_size, "Number of descriptors in send queue");
68 module_param_named(recv_queue_size, ipoib_recvq_size, int, 0444);
69 MODULE_PARM_DESC(recv_queue_size, "Number of descriptors in receive queue");
71 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
72 int ipoib_debug_level;
74 module_param_named(debug_level, ipoib_debug_level, int, 0644);
75 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
78 struct ipoib_path_iter {
79 struct net_device *dev;
80 struct ipoib_path path;
83 static const u8 ipv4_bcast_addr[] = {
84 0x00, 0xff, 0xff, 0xff,
85 0xff, 0x12, 0x40, 0x1b, 0x00, 0x00, 0x00, 0x00,
86 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff
89 struct workqueue_struct *ipoib_workqueue;
91 struct ib_sa_client ipoib_sa_client;
93 static void ipoib_add_one(struct ib_device *device);
94 static void ipoib_remove_one(struct ib_device *device, void *client_data);
95 static void ipoib_neigh_reclaim(struct rcu_head *rp);
96 static struct net_device *ipoib_get_net_dev_by_params(
97 struct ib_device *dev, u8 port, u16 pkey,
98 const union ib_gid *gid, const struct sockaddr *addr,
100 static int ipoib_set_mac(struct net_device *dev, void *addr);
101 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
104 static struct ib_client ipoib_client = {
106 .add = ipoib_add_one,
107 .remove = ipoib_remove_one,
108 .get_net_dev_by_params = ipoib_get_net_dev_by_params,
111 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
112 static int ipoib_netdev_event(struct notifier_block *this,
113 unsigned long event, void *ptr)
115 struct netdev_notifier_info *ni = ptr;
116 struct net_device *dev = ni->dev;
118 if (dev->netdev_ops->ndo_open != ipoib_open)
122 case NETDEV_REGISTER:
123 ipoib_create_debug_files(dev);
125 case NETDEV_CHANGENAME:
126 ipoib_delete_debug_files(dev);
127 ipoib_create_debug_files(dev);
129 case NETDEV_UNREGISTER:
130 ipoib_delete_debug_files(dev);
138 int ipoib_open(struct net_device *dev)
140 struct ipoib_dev_priv *priv = ipoib_priv(dev);
142 ipoib_dbg(priv, "bringing up interface\n");
144 netif_carrier_off(dev);
146 set_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
148 priv->sm_fullmember_sendonly_support = false;
150 if (ipoib_ib_dev_open(dev)) {
151 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags))
156 ipoib_ib_dev_up(dev);
158 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
159 struct ipoib_dev_priv *cpriv;
161 /* Bring up any child interfaces too */
162 down_read(&priv->vlan_rwsem);
163 list_for_each_entry(cpriv, &priv->child_intfs, list) {
166 flags = cpriv->dev->flags;
170 dev_change_flags(cpriv->dev, flags | IFF_UP);
172 up_read(&priv->vlan_rwsem);
175 netif_start_queue(dev);
180 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
185 static int ipoib_stop(struct net_device *dev)
187 struct ipoib_dev_priv *priv = ipoib_priv(dev);
189 ipoib_dbg(priv, "stopping interface\n");
191 clear_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags);
193 netif_stop_queue(dev);
195 ipoib_ib_dev_down(dev);
196 ipoib_ib_dev_stop(dev);
198 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
199 struct ipoib_dev_priv *cpriv;
201 /* Bring down any child interfaces too */
202 down_read(&priv->vlan_rwsem);
203 list_for_each_entry(cpriv, &priv->child_intfs, list) {
206 flags = cpriv->dev->flags;
207 if (!(flags & IFF_UP))
210 dev_change_flags(cpriv->dev, flags & ~IFF_UP);
212 up_read(&priv->vlan_rwsem);
218 static netdev_features_t ipoib_fix_features(struct net_device *dev, netdev_features_t features)
220 struct ipoib_dev_priv *priv = ipoib_priv(dev);
222 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
223 features &= ~(NETIF_F_IP_CSUM | NETIF_F_TSO);
228 static int ipoib_change_mtu(struct net_device *dev, int new_mtu)
230 struct ipoib_dev_priv *priv = ipoib_priv(dev);
233 /* dev->mtu > 2K ==> connected mode */
234 if (ipoib_cm_admin_enabled(dev)) {
235 if (new_mtu > ipoib_cm_max_mtu(dev))
238 if (new_mtu > priv->mcast_mtu)
239 ipoib_warn(priv, "mtu > %d will cause multicast packet drops.\n",
246 if (new_mtu > IPOIB_UD_MTU(priv->max_ib_mtu))
249 priv->admin_mtu = new_mtu;
251 if (priv->mcast_mtu < priv->admin_mtu)
252 ipoib_dbg(priv, "MTU must be smaller than the underlying "
253 "link layer MTU - 4 (%u)\n", priv->mcast_mtu);
255 new_mtu = min(priv->mcast_mtu, priv->admin_mtu);
257 if (priv->rn_ops->ndo_change_mtu) {
258 bool carrier_status = netif_carrier_ok(dev);
260 netif_carrier_off(dev);
262 /* notify lower level on the real mtu */
263 ret = priv->rn_ops->ndo_change_mtu(dev, new_mtu);
266 netif_carrier_on(dev);
274 static void ipoib_get_stats(struct net_device *dev,
275 struct rtnl_link_stats64 *stats)
277 struct ipoib_dev_priv *priv = ipoib_priv(dev);
279 if (priv->rn_ops->ndo_get_stats64)
280 priv->rn_ops->ndo_get_stats64(dev, stats);
282 netdev_stats_to_stats64(stats, &dev->stats);
285 /* Called with an RCU read lock taken */
286 static bool ipoib_is_dev_match_addr_rcu(const struct sockaddr *addr,
287 struct net_device *dev)
289 struct net *net = dev_net(dev);
290 struct in_device *in_dev;
291 struct sockaddr_in *addr_in = (struct sockaddr_in *)addr;
292 struct sockaddr_in6 *addr_in6 = (struct sockaddr_in6 *)addr;
295 switch (addr->sa_family) {
297 in_dev = in_dev_get(dev);
301 ret_addr = inet_confirm_addr(net, in_dev, 0,
302 addr_in->sin_addr.s_addr,
310 if (IS_ENABLED(CONFIG_IPV6) &&
311 ipv6_chk_addr(net, &addr_in6->sin6_addr, dev, 1))
320 * Find the master net_device on top of the given net_device.
321 * @dev: base IPoIB net_device
323 * Returns the master net_device with a reference held, or the same net_device
324 * if no master exists.
326 static struct net_device *ipoib_get_master_net_dev(struct net_device *dev)
328 struct net_device *master;
331 master = netdev_master_upper_dev_get_rcu(dev);
343 struct ipoib_walk_data {
344 const struct sockaddr *addr;
345 struct net_device *result;
348 static int ipoib_upper_walk(struct net_device *upper, void *_data)
350 struct ipoib_walk_data *data = _data;
353 if (ipoib_is_dev_match_addr_rcu(data->addr, upper)) {
355 data->result = upper;
363 * Find a net_device matching the given address, which is an upper device of
364 * the given net_device.
365 * @addr: IP address to look for.
366 * @dev: base IPoIB net_device
368 * If found, returns the net_device with a reference held. Otherwise return
371 static struct net_device *ipoib_get_net_dev_match_addr(
372 const struct sockaddr *addr, struct net_device *dev)
374 struct ipoib_walk_data data = {
379 if (ipoib_is_dev_match_addr_rcu(addr, dev)) {
385 netdev_walk_all_upper_dev_rcu(dev, ipoib_upper_walk, &data);
391 /* returns the number of IPoIB netdevs on top a given ipoib device matching a
392 * pkey_index and address, if one exists.
394 * @found_net_dev: contains a matching net_device if the return value >= 1,
395 * with a reference held. */
396 static int ipoib_match_gid_pkey_addr(struct ipoib_dev_priv *priv,
397 const union ib_gid *gid,
399 const struct sockaddr *addr,
401 struct net_device **found_net_dev)
403 struct ipoib_dev_priv *child_priv;
404 struct net_device *net_dev = NULL;
407 if (priv->pkey_index == pkey_index &&
408 (!gid || !memcmp(gid, &priv->local_gid, sizeof(*gid)))) {
410 net_dev = ipoib_get_master_net_dev(priv->dev);
412 /* Verify the net_device matches the IP address, as
413 * IPoIB child devices currently share a GID. */
414 net_dev = ipoib_get_net_dev_match_addr(addr, priv->dev);
418 *found_net_dev = net_dev;
425 /* Check child interfaces */
426 down_read_nested(&priv->vlan_rwsem, nesting);
427 list_for_each_entry(child_priv, &priv->child_intfs, list) {
428 matches += ipoib_match_gid_pkey_addr(child_priv, gid,
435 up_read(&priv->vlan_rwsem);
440 /* Returns the number of matching net_devs found (between 0 and 2). Also
441 * return the matching net_device in the @net_dev parameter, holding a
442 * reference to the net_device, if the number of matches >= 1 */
443 static int __ipoib_get_net_dev_by_params(struct list_head *dev_list, u8 port,
445 const union ib_gid *gid,
446 const struct sockaddr *addr,
447 struct net_device **net_dev)
449 struct ipoib_dev_priv *priv;
454 list_for_each_entry(priv, dev_list, list) {
455 if (priv->port != port)
458 matches += ipoib_match_gid_pkey_addr(priv, gid, pkey_index,
467 static struct net_device *ipoib_get_net_dev_by_params(
468 struct ib_device *dev, u8 port, u16 pkey,
469 const union ib_gid *gid, const struct sockaddr *addr,
472 struct net_device *net_dev;
473 struct list_head *dev_list = client_data;
478 if (!rdma_protocol_ib(dev, port))
481 ret = ib_find_cached_pkey(dev, port, pkey, &pkey_index);
488 /* See if we can find a unique device matching the L2 parameters */
489 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
490 gid, NULL, &net_dev);
501 /* Couldn't find a unique device with L2 parameters only. Use L3
502 * address to uniquely match the net device */
503 matches = __ipoib_get_net_dev_by_params(dev_list, port, pkey_index,
504 gid, addr, &net_dev);
509 dev_warn_ratelimited(&dev->dev,
510 "duplicate IP address detected\n");
517 int ipoib_set_mode(struct net_device *dev, const char *buf)
519 struct ipoib_dev_priv *priv = ipoib_priv(dev);
521 if ((test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
522 !strcmp(buf, "connected\n")) ||
523 (!test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags) &&
524 !strcmp(buf, "datagram\n"))) {
528 /* flush paths if we switch modes so that connections are restarted */
529 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
530 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
531 ipoib_warn(priv, "enabling connected mode "
532 "will cause multicast packet drops\n");
533 netdev_update_features(dev);
534 dev_set_mtu(dev, ipoib_cm_max_mtu(dev));
536 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM;
538 ipoib_flush_paths(dev);
539 return (!rtnl_trylock()) ? -EBUSY : 0;
542 if (!strcmp(buf, "datagram\n")) {
543 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
544 netdev_update_features(dev);
545 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
547 ipoib_flush_paths(dev);
548 return (!rtnl_trylock()) ? -EBUSY : 0;
554 struct ipoib_path *__path_find(struct net_device *dev, void *gid)
556 struct ipoib_dev_priv *priv = ipoib_priv(dev);
557 struct rb_node *n = priv->path_tree.rb_node;
558 struct ipoib_path *path;
562 path = rb_entry(n, struct ipoib_path, rb_node);
564 ret = memcmp(gid, path->pathrec.dgid.raw,
565 sizeof (union ib_gid));
578 static int __path_add(struct net_device *dev, struct ipoib_path *path)
580 struct ipoib_dev_priv *priv = ipoib_priv(dev);
581 struct rb_node **n = &priv->path_tree.rb_node;
582 struct rb_node *pn = NULL;
583 struct ipoib_path *tpath;
588 tpath = rb_entry(pn, struct ipoib_path, rb_node);
590 ret = memcmp(path->pathrec.dgid.raw, tpath->pathrec.dgid.raw,
591 sizeof (union ib_gid));
600 rb_link_node(&path->rb_node, pn, n);
601 rb_insert_color(&path->rb_node, &priv->path_tree);
603 list_add_tail(&path->list, &priv->path_list);
608 static void path_free(struct net_device *dev, struct ipoib_path *path)
612 while ((skb = __skb_dequeue(&path->queue)))
613 dev_kfree_skb_irq(skb);
615 ipoib_dbg(ipoib_priv(dev), "path_free\n");
617 /* remove all neigh connected to this path */
618 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
621 ipoib_put_ah(path->ah);
626 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
628 struct ipoib_path_iter *ipoib_path_iter_init(struct net_device *dev)
630 struct ipoib_path_iter *iter;
632 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
637 memset(iter->path.pathrec.dgid.raw, 0, 16);
639 if (ipoib_path_iter_next(iter)) {
647 int ipoib_path_iter_next(struct ipoib_path_iter *iter)
649 struct ipoib_dev_priv *priv = ipoib_priv(iter->dev);
651 struct ipoib_path *path;
654 spin_lock_irq(&priv->lock);
656 n = rb_first(&priv->path_tree);
659 path = rb_entry(n, struct ipoib_path, rb_node);
661 if (memcmp(iter->path.pathrec.dgid.raw, path->pathrec.dgid.raw,
662 sizeof (union ib_gid)) < 0) {
671 spin_unlock_irq(&priv->lock);
676 void ipoib_path_iter_read(struct ipoib_path_iter *iter,
677 struct ipoib_path *path)
682 #endif /* CONFIG_INFINIBAND_IPOIB_DEBUG */
684 void ipoib_mark_paths_invalid(struct net_device *dev)
686 struct ipoib_dev_priv *priv = ipoib_priv(dev);
687 struct ipoib_path *path, *tp;
689 spin_lock_irq(&priv->lock);
691 list_for_each_entry_safe(path, tp, &priv->path_list, list) {
692 ipoib_dbg(priv, "mark path LID 0x%08x GID %pI6 invalid\n",
693 be32_to_cpu(sa_path_get_dlid(&path->pathrec)),
694 path->pathrec.dgid.raw);
699 spin_unlock_irq(&priv->lock);
702 static void push_pseudo_header(struct sk_buff *skb, const char *daddr)
704 struct ipoib_pseudo_header *phdr;
706 phdr = skb_push(skb, sizeof(*phdr));
707 memcpy(phdr->hwaddr, daddr, INFINIBAND_ALEN);
710 void ipoib_flush_paths(struct net_device *dev)
712 struct ipoib_dev_priv *priv = ipoib_priv(dev);
713 struct ipoib_path *path, *tp;
714 LIST_HEAD(remove_list);
717 netif_tx_lock_bh(dev);
718 spin_lock_irqsave(&priv->lock, flags);
720 list_splice_init(&priv->path_list, &remove_list);
722 list_for_each_entry(path, &remove_list, list)
723 rb_erase(&path->rb_node, &priv->path_tree);
725 list_for_each_entry_safe(path, tp, &remove_list, list) {
727 ib_sa_cancel_query(path->query_id, path->query);
728 spin_unlock_irqrestore(&priv->lock, flags);
729 netif_tx_unlock_bh(dev);
730 wait_for_completion(&path->done);
731 path_free(dev, path);
732 netif_tx_lock_bh(dev);
733 spin_lock_irqsave(&priv->lock, flags);
736 spin_unlock_irqrestore(&priv->lock, flags);
737 netif_tx_unlock_bh(dev);
740 static void path_rec_completion(int status,
741 struct sa_path_rec *pathrec,
744 struct ipoib_path *path = path_ptr;
745 struct net_device *dev = path->dev;
746 struct ipoib_dev_priv *priv = ipoib_priv(dev);
747 struct ipoib_ah *ah = NULL;
748 struct ipoib_ah *old_ah = NULL;
749 struct ipoib_neigh *neigh, *tn;
750 struct sk_buff_head skqueue;
755 ipoib_dbg(priv, "PathRec LID 0x%04x for GID %pI6\n",
756 be32_to_cpu(sa_path_get_dlid(pathrec)),
759 ipoib_dbg(priv, "PathRec status %d for GID %pI6\n",
760 status, path->pathrec.dgid.raw);
762 skb_queue_head_init(&skqueue);
765 struct rdma_ah_attr av;
767 if (!ib_init_ah_attr_from_path(priv->ca, priv->port,
768 pathrec, &av, NULL)) {
769 ah = ipoib_create_ah(dev, priv->pd, &av);
770 rdma_destroy_ah_attr(&av);
774 spin_lock_irqsave(&priv->lock, flags);
776 if (!IS_ERR_OR_NULL(ah)) {
778 * pathrec.dgid is used as the database key from the LLADDR,
779 * it must remain unchanged even if the SA returns a different
780 * GID to use in the AH.
782 if (memcmp(pathrec->dgid.raw, path->pathrec.dgid.raw,
783 sizeof(union ib_gid))) {
786 "%s got PathRec for gid %pI6 while asked for %pI6\n",
787 dev->name, pathrec->dgid.raw,
788 path->pathrec.dgid.raw);
789 memcpy(pathrec->dgid.raw, path->pathrec.dgid.raw,
790 sizeof(union ib_gid));
793 path->pathrec = *pathrec;
798 ipoib_dbg(priv, "created address handle %p for LID 0x%04x, SL %d\n",
799 ah, be32_to_cpu(sa_path_get_dlid(pathrec)),
802 while ((skb = __skb_dequeue(&path->queue)))
803 __skb_queue_tail(&skqueue, skb);
805 list_for_each_entry_safe(neigh, tn, &path->neigh_list, list) {
807 WARN_ON(neigh->ah != old_ah);
809 * Dropping the ah reference inside
810 * priv->lock is safe here, because we
811 * will hold one more reference from
812 * the original value of path->ah (ie
815 ipoib_put_ah(neigh->ah);
817 kref_get(&path->ah->ref);
818 neigh->ah = path->ah;
820 if (ipoib_cm_enabled(dev, neigh->daddr)) {
821 if (!ipoib_cm_get(neigh))
822 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev,
825 if (!ipoib_cm_get(neigh)) {
826 ipoib_neigh_free(neigh);
831 while ((skb = __skb_dequeue(&neigh->queue)))
832 __skb_queue_tail(&skqueue, skb);
838 complete(&path->done);
840 spin_unlock_irqrestore(&priv->lock, flags);
842 if (IS_ERR_OR_NULL(ah))
843 ipoib_del_neighs_by_gid(dev, path->pathrec.dgid.raw);
846 ipoib_put_ah(old_ah);
848 while ((skb = __skb_dequeue(&skqueue))) {
851 ret = dev_queue_xmit(skb);
853 ipoib_warn(priv, "%s: dev_queue_xmit failed to re-queue packet, ret:%d\n",
858 static void init_path_rec(struct ipoib_dev_priv *priv, struct ipoib_path *path,
861 path->dev = priv->dev;
863 if (rdma_cap_opa_ah(priv->ca, priv->port))
864 path->pathrec.rec_type = SA_PATH_REC_TYPE_OPA;
866 path->pathrec.rec_type = SA_PATH_REC_TYPE_IB;
868 memcpy(path->pathrec.dgid.raw, gid, sizeof(union ib_gid));
869 path->pathrec.sgid = priv->local_gid;
870 path->pathrec.pkey = cpu_to_be16(priv->pkey);
871 path->pathrec.numb_path = 1;
872 path->pathrec.traffic_class = priv->broadcast->mcmember.traffic_class;
875 static struct ipoib_path *path_rec_create(struct net_device *dev, void *gid)
877 struct ipoib_dev_priv *priv = ipoib_priv(dev);
878 struct ipoib_path *path;
880 if (!priv->broadcast)
883 path = kzalloc(sizeof(*path), GFP_ATOMIC);
887 skb_queue_head_init(&path->queue);
889 INIT_LIST_HEAD(&path->neigh_list);
891 init_path_rec(priv, path, gid);
896 static int path_rec_start(struct net_device *dev,
897 struct ipoib_path *path)
899 struct ipoib_dev_priv *priv = ipoib_priv(dev);
901 ipoib_dbg(priv, "Start path record lookup for %pI6\n",
902 path->pathrec.dgid.raw);
904 init_completion(&path->done);
907 ib_sa_path_rec_get(&ipoib_sa_client, priv->ca, priv->port,
909 IB_SA_PATH_REC_DGID |
910 IB_SA_PATH_REC_SGID |
911 IB_SA_PATH_REC_NUMB_PATH |
912 IB_SA_PATH_REC_TRAFFIC_CLASS |
917 if (path->query_id < 0) {
918 ipoib_warn(priv, "ib_sa_path_rec_get failed: %d\n", path->query_id);
920 complete(&path->done);
921 return path->query_id;
927 static void neigh_refresh_path(struct ipoib_neigh *neigh, u8 *daddr,
928 struct net_device *dev)
930 struct ipoib_dev_priv *priv = ipoib_priv(dev);
931 struct ipoib_path *path;
934 spin_lock_irqsave(&priv->lock, flags);
936 path = __path_find(dev, daddr + 4);
940 path_rec_start(dev, path);
942 spin_unlock_irqrestore(&priv->lock, flags);
945 static struct ipoib_neigh *neigh_add_path(struct sk_buff *skb, u8 *daddr,
946 struct net_device *dev)
948 struct ipoib_dev_priv *priv = ipoib_priv(dev);
949 struct rdma_netdev *rn = netdev_priv(dev);
950 struct ipoib_path *path;
951 struct ipoib_neigh *neigh;
954 spin_lock_irqsave(&priv->lock, flags);
955 neigh = ipoib_neigh_alloc(daddr, dev);
957 spin_unlock_irqrestore(&priv->lock, flags);
958 ++dev->stats.tx_dropped;
959 dev_kfree_skb_any(skb);
963 /* To avoid race condition, make sure that the
964 * neigh will be added only once.
966 if (unlikely(!list_empty(&neigh->list))) {
967 spin_unlock_irqrestore(&priv->lock, flags);
971 path = __path_find(dev, daddr + 4);
973 path = path_rec_create(dev, daddr + 4);
977 __path_add(dev, path);
980 list_add_tail(&neigh->list, &path->neigh_list);
982 if (path->ah && path->ah->valid) {
983 kref_get(&path->ah->ref);
984 neigh->ah = path->ah;
986 if (ipoib_cm_enabled(dev, neigh->daddr)) {
987 if (!ipoib_cm_get(neigh))
988 ipoib_cm_set(neigh, ipoib_cm_create_tx(dev, path, neigh));
989 if (!ipoib_cm_get(neigh)) {
990 ipoib_neigh_free(neigh);
993 if (skb_queue_len(&neigh->queue) <
994 IPOIB_MAX_PATH_REC_QUEUE) {
995 push_pseudo_header(skb, neigh->daddr);
996 __skb_queue_tail(&neigh->queue, skb);
998 ipoib_warn(priv, "queue length limit %d. Packet drop.\n",
999 skb_queue_len(&neigh->queue));
1003 spin_unlock_irqrestore(&priv->lock, flags);
1004 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1006 ipoib_neigh_put(neigh);
1012 if (!path->query && path_rec_start(dev, path))
1014 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1015 push_pseudo_header(skb, neigh->daddr);
1016 __skb_queue_tail(&neigh->queue, skb);
1022 spin_unlock_irqrestore(&priv->lock, flags);
1023 ipoib_neigh_put(neigh);
1027 ipoib_neigh_free(neigh);
1029 ++dev->stats.tx_dropped;
1030 dev_kfree_skb_any(skb);
1032 spin_unlock_irqrestore(&priv->lock, flags);
1033 ipoib_neigh_put(neigh);
1038 static void unicast_arp_send(struct sk_buff *skb, struct net_device *dev,
1039 struct ipoib_pseudo_header *phdr)
1041 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1042 struct rdma_netdev *rn = netdev_priv(dev);
1043 struct ipoib_path *path;
1044 unsigned long flags;
1046 spin_lock_irqsave(&priv->lock, flags);
1048 /* no broadcast means that all paths are (going to be) not valid */
1049 if (!priv->broadcast)
1050 goto drop_and_unlock;
1052 path = __path_find(dev, phdr->hwaddr + 4);
1053 if (!path || !path->ah || !path->ah->valid) {
1055 path = path_rec_create(dev, phdr->hwaddr + 4);
1057 goto drop_and_unlock;
1058 __path_add(dev, path);
1061 * make sure there are no changes in the existing
1064 init_path_rec(priv, path, phdr->hwaddr + 4);
1066 if (!path->query && path_rec_start(dev, path)) {
1067 goto drop_and_unlock;
1070 if (skb_queue_len(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1071 push_pseudo_header(skb, phdr->hwaddr);
1072 __skb_queue_tail(&path->queue, skb);
1075 goto drop_and_unlock;
1079 spin_unlock_irqrestore(&priv->lock, flags);
1080 ipoib_dbg(priv, "Send unicast ARP to %08x\n",
1081 be32_to_cpu(sa_path_get_dlid(&path->pathrec)));
1082 path->ah->last_send = rn->send(dev, skb, path->ah->ah,
1083 IPOIB_QPN(phdr->hwaddr));
1087 ++dev->stats.tx_dropped;
1088 dev_kfree_skb_any(skb);
1090 spin_unlock_irqrestore(&priv->lock, flags);
1093 static netdev_tx_t ipoib_start_xmit(struct sk_buff *skb, struct net_device *dev)
1095 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1096 struct rdma_netdev *rn = netdev_priv(dev);
1097 struct ipoib_neigh *neigh;
1098 struct ipoib_pseudo_header *phdr;
1099 struct ipoib_header *header;
1100 unsigned long flags;
1102 phdr = (struct ipoib_pseudo_header *) skb->data;
1103 skb_pull(skb, sizeof(*phdr));
1104 header = (struct ipoib_header *) skb->data;
1106 if (unlikely(phdr->hwaddr[4] == 0xff)) {
1107 /* multicast, arrange "if" according to probability */
1108 if ((header->proto != htons(ETH_P_IP)) &&
1109 (header->proto != htons(ETH_P_IPV6)) &&
1110 (header->proto != htons(ETH_P_ARP)) &&
1111 (header->proto != htons(ETH_P_RARP)) &&
1112 (header->proto != htons(ETH_P_TIPC))) {
1113 /* ethertype not supported by IPoIB */
1114 ++dev->stats.tx_dropped;
1115 dev_kfree_skb_any(skb);
1116 return NETDEV_TX_OK;
1118 /* Add in the P_Key for multicast*/
1119 phdr->hwaddr[8] = (priv->pkey >> 8) & 0xff;
1120 phdr->hwaddr[9] = priv->pkey & 0xff;
1122 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1124 goto send_using_neigh;
1125 ipoib_mcast_send(dev, phdr->hwaddr, skb);
1126 return NETDEV_TX_OK;
1129 /* unicast, arrange "switch" according to probability */
1130 switch (header->proto) {
1131 case htons(ETH_P_IP):
1132 case htons(ETH_P_IPV6):
1133 case htons(ETH_P_TIPC):
1134 neigh = ipoib_neigh_get(dev, phdr->hwaddr);
1135 if (unlikely(!neigh)) {
1136 neigh = neigh_add_path(skb, phdr->hwaddr, dev);
1138 return NETDEV_TX_OK;
1141 case htons(ETH_P_ARP):
1142 case htons(ETH_P_RARP):
1143 /* for unicast ARP and RARP should always perform path find */
1144 unicast_arp_send(skb, dev, phdr);
1145 return NETDEV_TX_OK;
1147 /* ethertype not supported by IPoIB */
1148 ++dev->stats.tx_dropped;
1149 dev_kfree_skb_any(skb);
1150 return NETDEV_TX_OK;
1154 /* note we now hold a ref to neigh */
1155 if (ipoib_cm_get(neigh)) {
1156 if (ipoib_cm_up(neigh)) {
1157 ipoib_cm_send(dev, skb, ipoib_cm_get(neigh));
1160 } else if (neigh->ah && neigh->ah->valid) {
1161 neigh->ah->last_send = rn->send(dev, skb, neigh->ah->ah,
1162 IPOIB_QPN(phdr->hwaddr));
1164 } else if (neigh->ah) {
1165 neigh_refresh_path(neigh, phdr->hwaddr, dev);
1168 if (skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE) {
1169 push_pseudo_header(skb, phdr->hwaddr);
1170 spin_lock_irqsave(&priv->lock, flags);
1171 __skb_queue_tail(&neigh->queue, skb);
1172 spin_unlock_irqrestore(&priv->lock, flags);
1174 ++dev->stats.tx_dropped;
1175 dev_kfree_skb_any(skb);
1179 ipoib_neigh_put(neigh);
1181 return NETDEV_TX_OK;
1184 static void ipoib_timeout(struct net_device *dev)
1186 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1188 ipoib_warn(priv, "transmit timeout: latency %d msecs\n",
1189 jiffies_to_msecs(jiffies - dev_trans_start(dev)));
1190 ipoib_warn(priv, "queue stopped %d, tx_head %u, tx_tail %u\n",
1191 netif_queue_stopped(dev),
1192 priv->tx_head, priv->tx_tail);
1193 /* XXX reset QP, etc. */
1196 static int ipoib_hard_header(struct sk_buff *skb,
1197 struct net_device *dev,
1198 unsigned short type,
1203 struct ipoib_header *header;
1205 header = skb_push(skb, sizeof(*header));
1207 header->proto = htons(type);
1208 header->reserved = 0;
1211 * we don't rely on dst_entry structure, always stuff the
1212 * destination address into skb hard header so we can figure out where
1213 * to send the packet later.
1215 push_pseudo_header(skb, daddr);
1217 return IPOIB_HARD_LEN;
1220 static void ipoib_set_mcast_list(struct net_device *dev)
1222 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1224 if (!test_bit(IPOIB_FLAG_OPER_UP, &priv->flags)) {
1225 ipoib_dbg(priv, "IPOIB_FLAG_OPER_UP not set");
1229 queue_work(priv->wq, &priv->restart_task);
1232 static int ipoib_get_iflink(const struct net_device *dev)
1234 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1236 /* parent interface */
1237 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags))
1238 return dev->ifindex;
1240 /* child/vlan interface */
1241 return priv->parent->ifindex;
1244 static u32 ipoib_addr_hash(struct ipoib_neigh_hash *htbl, u8 *daddr)
1247 * Use only the address parts that contributes to spreading
1248 * The subnet prefix is not used as one can not connect to
1249 * same remote port (GUID) using the same remote QPN via two
1250 * different subnets.
1252 /* qpn octets[1:4) & port GUID octets[12:20) */
1253 u32 *d32 = (u32 *) daddr;
1256 hv = jhash_3words(d32[3], d32[4], IPOIB_QPN_MASK & d32[0], 0);
1257 return hv & htbl->mask;
1260 struct ipoib_neigh *ipoib_neigh_get(struct net_device *dev, u8 *daddr)
1262 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1263 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1264 struct ipoib_neigh_hash *htbl;
1265 struct ipoib_neigh *neigh = NULL;
1270 htbl = rcu_dereference_bh(ntbl->htbl);
1275 hash_val = ipoib_addr_hash(htbl, daddr);
1276 for (neigh = rcu_dereference_bh(htbl->buckets[hash_val]);
1278 neigh = rcu_dereference_bh(neigh->hnext)) {
1279 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1280 /* found, take one ref on behalf of the caller */
1281 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1287 if (likely(skb_queue_len(&neigh->queue) < IPOIB_MAX_PATH_REC_QUEUE))
1288 neigh->alive = jiffies;
1294 rcu_read_unlock_bh();
1298 static void __ipoib_reap_neigh(struct ipoib_dev_priv *priv)
1300 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1301 struct ipoib_neigh_hash *htbl;
1302 unsigned long neigh_obsolete;
1304 unsigned long flags;
1306 LIST_HEAD(remove_list);
1308 spin_lock_irqsave(&priv->lock, flags);
1310 htbl = rcu_dereference_protected(ntbl->htbl,
1311 lockdep_is_held(&priv->lock));
1316 /* neigh is obsolete if it was idle for two GC periods */
1317 dt = 2 * arp_tbl.gc_interval;
1318 neigh_obsolete = jiffies - dt;
1320 for (i = 0; i < htbl->size; i++) {
1321 struct ipoib_neigh *neigh;
1322 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1324 while ((neigh = rcu_dereference_protected(*np,
1325 lockdep_is_held(&priv->lock))) != NULL) {
1326 /* was the neigh idle for two GC periods */
1327 if (time_after(neigh_obsolete, neigh->alive)) {
1329 ipoib_check_and_add_mcast_sendonly(priv, neigh->daddr + 4, &remove_list);
1331 rcu_assign_pointer(*np,
1332 rcu_dereference_protected(neigh->hnext,
1333 lockdep_is_held(&priv->lock)));
1334 /* remove from path/mc list */
1335 list_del_init(&neigh->list);
1336 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1345 spin_unlock_irqrestore(&priv->lock, flags);
1346 ipoib_mcast_remove_list(&remove_list);
1349 static void ipoib_reap_neigh(struct work_struct *work)
1351 struct ipoib_dev_priv *priv =
1352 container_of(work, struct ipoib_dev_priv, neigh_reap_task.work);
1354 __ipoib_reap_neigh(priv);
1356 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1357 arp_tbl.gc_interval);
1361 static struct ipoib_neigh *ipoib_neigh_ctor(u8 *daddr,
1362 struct net_device *dev)
1364 struct ipoib_neigh *neigh;
1366 neigh = kzalloc(sizeof(*neigh), GFP_ATOMIC);
1371 memcpy(&neigh->daddr, daddr, sizeof(neigh->daddr));
1372 skb_queue_head_init(&neigh->queue);
1373 INIT_LIST_HEAD(&neigh->list);
1374 ipoib_cm_set(neigh, NULL);
1375 /* one ref on behalf of the caller */
1376 atomic_set(&neigh->refcnt, 1);
1381 struct ipoib_neigh *ipoib_neigh_alloc(u8 *daddr,
1382 struct net_device *dev)
1384 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1385 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1386 struct ipoib_neigh_hash *htbl;
1387 struct ipoib_neigh *neigh;
1390 htbl = rcu_dereference_protected(ntbl->htbl,
1391 lockdep_is_held(&priv->lock));
1397 /* need to add a new neigh, but maybe some other thread succeeded?
1398 * recalc hash, maybe hash resize took place so we do a search
1400 hash_val = ipoib_addr_hash(htbl, daddr);
1401 for (neigh = rcu_dereference_protected(htbl->buckets[hash_val],
1402 lockdep_is_held(&priv->lock));
1404 neigh = rcu_dereference_protected(neigh->hnext,
1405 lockdep_is_held(&priv->lock))) {
1406 if (memcmp(daddr, neigh->daddr, INFINIBAND_ALEN) == 0) {
1407 /* found, take one ref on behalf of the caller */
1408 if (!atomic_inc_not_zero(&neigh->refcnt)) {
1413 neigh->alive = jiffies;
1418 neigh = ipoib_neigh_ctor(daddr, dev);
1422 /* one ref on behalf of the hash table */
1423 atomic_inc(&neigh->refcnt);
1424 neigh->alive = jiffies;
1426 rcu_assign_pointer(neigh->hnext,
1427 rcu_dereference_protected(htbl->buckets[hash_val],
1428 lockdep_is_held(&priv->lock)));
1429 rcu_assign_pointer(htbl->buckets[hash_val], neigh);
1430 atomic_inc(&ntbl->entries);
1437 void ipoib_neigh_dtor(struct ipoib_neigh *neigh)
1439 /* neigh reference count was dropprd to zero */
1440 struct net_device *dev = neigh->dev;
1441 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1442 struct sk_buff *skb;
1444 ipoib_put_ah(neigh->ah);
1445 while ((skb = __skb_dequeue(&neigh->queue))) {
1446 ++dev->stats.tx_dropped;
1447 dev_kfree_skb_any(skb);
1449 if (ipoib_cm_get(neigh))
1450 ipoib_cm_destroy_tx(ipoib_cm_get(neigh));
1451 ipoib_dbg(ipoib_priv(dev),
1452 "neigh free for %06x %pI6\n",
1453 IPOIB_QPN(neigh->daddr),
1456 if (atomic_dec_and_test(&priv->ntbl.entries)) {
1457 if (test_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags))
1458 complete(&priv->ntbl.flushed);
1462 static void ipoib_neigh_reclaim(struct rcu_head *rp)
1464 /* Called as a result of removal from hash table */
1465 struct ipoib_neigh *neigh = container_of(rp, struct ipoib_neigh, rcu);
1466 /* note TX context may hold another ref */
1467 ipoib_neigh_put(neigh);
1470 void ipoib_neigh_free(struct ipoib_neigh *neigh)
1472 struct net_device *dev = neigh->dev;
1473 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1474 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1475 struct ipoib_neigh_hash *htbl;
1476 struct ipoib_neigh __rcu **np;
1477 struct ipoib_neigh *n;
1480 htbl = rcu_dereference_protected(ntbl->htbl,
1481 lockdep_is_held(&priv->lock));
1485 hash_val = ipoib_addr_hash(htbl, neigh->daddr);
1486 np = &htbl->buckets[hash_val];
1487 for (n = rcu_dereference_protected(*np,
1488 lockdep_is_held(&priv->lock));
1490 n = rcu_dereference_protected(*np,
1491 lockdep_is_held(&priv->lock))) {
1494 rcu_assign_pointer(*np,
1495 rcu_dereference_protected(neigh->hnext,
1496 lockdep_is_held(&priv->lock)));
1497 /* remove from parent list */
1498 list_del_init(&neigh->list);
1499 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1507 static int ipoib_neigh_hash_init(struct ipoib_dev_priv *priv)
1509 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1510 struct ipoib_neigh_hash *htbl;
1511 struct ipoib_neigh __rcu **buckets;
1514 clear_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1516 htbl = kzalloc(sizeof(*htbl), GFP_KERNEL);
1519 size = roundup_pow_of_two(arp_tbl.gc_thresh3);
1520 buckets = kvcalloc(size, sizeof(*buckets), GFP_KERNEL);
1526 htbl->mask = (size - 1);
1527 htbl->buckets = buckets;
1528 RCU_INIT_POINTER(ntbl->htbl, htbl);
1530 atomic_set(&ntbl->entries, 0);
1532 /* start garbage collection */
1533 queue_delayed_work(priv->wq, &priv->neigh_reap_task,
1534 arp_tbl.gc_interval);
1539 static void neigh_hash_free_rcu(struct rcu_head *head)
1541 struct ipoib_neigh_hash *htbl = container_of(head,
1542 struct ipoib_neigh_hash,
1544 struct ipoib_neigh __rcu **buckets = htbl->buckets;
1545 struct ipoib_neigh_table *ntbl = htbl->ntbl;
1549 complete(&ntbl->deleted);
1552 void ipoib_del_neighs_by_gid(struct net_device *dev, u8 *gid)
1554 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1555 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1556 struct ipoib_neigh_hash *htbl;
1557 unsigned long flags;
1560 /* remove all neigh connected to a given path or mcast */
1561 spin_lock_irqsave(&priv->lock, flags);
1563 htbl = rcu_dereference_protected(ntbl->htbl,
1564 lockdep_is_held(&priv->lock));
1569 for (i = 0; i < htbl->size; i++) {
1570 struct ipoib_neigh *neigh;
1571 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1573 while ((neigh = rcu_dereference_protected(*np,
1574 lockdep_is_held(&priv->lock))) != NULL) {
1575 /* delete neighs belong to this parent */
1576 if (!memcmp(gid, neigh->daddr + 4, sizeof (union ib_gid))) {
1577 rcu_assign_pointer(*np,
1578 rcu_dereference_protected(neigh->hnext,
1579 lockdep_is_held(&priv->lock)));
1580 /* remove from parent list */
1581 list_del_init(&neigh->list);
1582 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1590 spin_unlock_irqrestore(&priv->lock, flags);
1593 static void ipoib_flush_neighs(struct ipoib_dev_priv *priv)
1595 struct ipoib_neigh_table *ntbl = &priv->ntbl;
1596 struct ipoib_neigh_hash *htbl;
1597 unsigned long flags;
1598 int i, wait_flushed = 0;
1600 init_completion(&priv->ntbl.flushed);
1601 set_bit(IPOIB_NEIGH_TBL_FLUSH, &priv->flags);
1603 spin_lock_irqsave(&priv->lock, flags);
1605 htbl = rcu_dereference_protected(ntbl->htbl,
1606 lockdep_is_held(&priv->lock));
1610 wait_flushed = atomic_read(&priv->ntbl.entries);
1614 for (i = 0; i < htbl->size; i++) {
1615 struct ipoib_neigh *neigh;
1616 struct ipoib_neigh __rcu **np = &htbl->buckets[i];
1618 while ((neigh = rcu_dereference_protected(*np,
1619 lockdep_is_held(&priv->lock))) != NULL) {
1620 rcu_assign_pointer(*np,
1621 rcu_dereference_protected(neigh->hnext,
1622 lockdep_is_held(&priv->lock)));
1623 /* remove from path/mc list */
1624 list_del_init(&neigh->list);
1625 call_rcu(&neigh->rcu, ipoib_neigh_reclaim);
1630 rcu_assign_pointer(ntbl->htbl, NULL);
1631 call_rcu(&htbl->rcu, neigh_hash_free_rcu);
1634 spin_unlock_irqrestore(&priv->lock, flags);
1636 wait_for_completion(&priv->ntbl.flushed);
1639 static void ipoib_neigh_hash_uninit(struct net_device *dev)
1641 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1643 ipoib_dbg(priv, "ipoib_neigh_hash_uninit\n");
1644 init_completion(&priv->ntbl.deleted);
1646 cancel_delayed_work_sync(&priv->neigh_reap_task);
1648 ipoib_flush_neighs(priv);
1650 wait_for_completion(&priv->ntbl.deleted);
1653 static void ipoib_napi_add(struct net_device *dev)
1655 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1657 netif_napi_add(dev, &priv->recv_napi, ipoib_rx_poll, IPOIB_NUM_WC);
1658 netif_napi_add(dev, &priv->send_napi, ipoib_tx_poll, MAX_SEND_CQE);
1661 static void ipoib_napi_del(struct net_device *dev)
1663 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1665 netif_napi_del(&priv->recv_napi);
1666 netif_napi_del(&priv->send_napi);
1669 static void ipoib_dev_uninit_default(struct net_device *dev)
1671 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1673 ipoib_transport_dev_cleanup(dev);
1675 ipoib_napi_del(dev);
1677 ipoib_cm_dev_cleanup(dev);
1679 kfree(priv->rx_ring);
1680 vfree(priv->tx_ring);
1682 priv->rx_ring = NULL;
1683 priv->tx_ring = NULL;
1686 static int ipoib_dev_init_default(struct net_device *dev)
1688 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1690 ipoib_napi_add(dev);
1692 /* Allocate RX/TX "rings" to hold queued skbs */
1693 priv->rx_ring = kcalloc(ipoib_recvq_size,
1694 sizeof(*priv->rx_ring),
1699 priv->tx_ring = vzalloc(array_size(ipoib_sendq_size,
1700 sizeof(*priv->tx_ring)));
1701 if (!priv->tx_ring) {
1702 pr_warn("%s: failed to allocate TX ring (%d entries)\n",
1703 priv->ca->name, ipoib_sendq_size);
1704 goto out_rx_ring_cleanup;
1707 /* priv->tx_head, tx_tail & tx_outstanding are already 0 */
1709 if (ipoib_transport_dev_init(dev, priv->ca)) {
1710 pr_warn("%s: ipoib_transport_dev_init failed\n",
1712 goto out_tx_ring_cleanup;
1715 /* after qp created set dev address */
1716 priv->dev->dev_addr[1] = (priv->qp->qp_num >> 16) & 0xff;
1717 priv->dev->dev_addr[2] = (priv->qp->qp_num >> 8) & 0xff;
1718 priv->dev->dev_addr[3] = (priv->qp->qp_num) & 0xff;
1722 out_tx_ring_cleanup:
1723 vfree(priv->tx_ring);
1725 out_rx_ring_cleanup:
1726 kfree(priv->rx_ring);
1729 ipoib_napi_del(dev);
1733 static int ipoib_ioctl(struct net_device *dev, struct ifreq *ifr,
1736 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1738 if (!priv->rn_ops->ndo_do_ioctl)
1741 return priv->rn_ops->ndo_do_ioctl(dev, ifr, cmd);
1744 static int ipoib_dev_init(struct net_device *dev)
1746 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1752 * the various IPoIB tasks assume they will never race against
1753 * themselves, so always use a single thread workqueue
1755 priv->wq = alloc_ordered_workqueue("ipoib_wq", WQ_MEM_RECLAIM);
1757 pr_warn("%s: failed to allocate device WQ\n", dev->name);
1761 /* create pd, which used both for control and datapath*/
1762 priv->pd = ib_alloc_pd(priv->ca, 0);
1763 if (IS_ERR(priv->pd)) {
1764 pr_warn("%s: failed to allocate PD\n", priv->ca->name);
1768 ret = priv->rn_ops->ndo_init(dev);
1770 pr_warn("%s failed to init HW resource\n", dev->name);
1774 ret = ipoib_neigh_hash_init(priv);
1776 pr_warn("%s failed to init neigh hash\n", dev->name);
1777 goto out_dev_uninit;
1780 if (dev->flags & IFF_UP) {
1781 if (ipoib_ib_dev_open(dev)) {
1782 pr_warn("%s failed to open device\n", dev->name);
1784 goto out_hash_uninit;
1791 ipoib_neigh_hash_uninit(dev);
1794 ipoib_ib_dev_cleanup(dev);
1798 ib_dealloc_pd(priv->pd);
1804 destroy_workqueue(priv->wq);
1813 * This must be called before doing an unregister_netdev on a parent device to
1814 * shutdown the IB event handler.
1816 static void ipoib_parent_unregister_pre(struct net_device *ndev)
1818 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1821 * ipoib_set_mac checks netif_running before pushing work, clearing
1822 * running ensures the it will not add more work.
1825 dev_change_flags(priv->dev, priv->dev->flags & ~IFF_UP);
1828 /* ipoib_event() cannot be running once this returns */
1829 ib_unregister_event_handler(&priv->event_handler);
1832 * Work on the queue grabs the rtnl lock, so this cannot be done while
1835 flush_workqueue(ipoib_workqueue);
1838 static void ipoib_set_dev_features(struct ipoib_dev_priv *priv)
1840 priv->hca_caps = priv->ca->attrs.device_cap_flags;
1842 if (priv->hca_caps & IB_DEVICE_UD_IP_CSUM) {
1843 priv->dev->hw_features |= NETIF_F_IP_CSUM | NETIF_F_RXCSUM;
1845 if (priv->hca_caps & IB_DEVICE_UD_TSO)
1846 priv->dev->hw_features |= NETIF_F_TSO;
1848 priv->dev->features |= priv->dev->hw_features;
1852 static int ipoib_parent_init(struct net_device *ndev)
1854 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1855 struct ib_port_attr attr;
1858 result = ib_query_port(priv->ca, priv->port, &attr);
1860 pr_warn("%s: ib_query_port %d failed\n", priv->ca->name,
1864 priv->max_ib_mtu = ib_mtu_enum_to_int(attr.max_mtu);
1866 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey);
1868 pr_warn("%s: ib_query_pkey port %d failed (ret = %d)\n",
1869 priv->ca->name, priv->port, result);
1873 result = rdma_query_gid(priv->ca, priv->port, 0, &priv->local_gid);
1875 pr_warn("%s: rdma_query_gid port %d failed (ret = %d)\n",
1876 priv->ca->name, priv->port, result);
1879 memcpy(priv->dev->dev_addr + 4, priv->local_gid.raw,
1880 sizeof(union ib_gid));
1882 SET_NETDEV_DEV(priv->dev, priv->ca->dev.parent);
1883 priv->dev->dev_id = priv->port - 1;
1888 static void ipoib_child_init(struct net_device *ndev)
1890 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1891 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1893 dev_hold(priv->parent);
1895 down_write(&ppriv->vlan_rwsem);
1896 list_add_tail(&priv->list, &ppriv->child_intfs);
1897 up_write(&ppriv->vlan_rwsem);
1899 priv->max_ib_mtu = ppriv->max_ib_mtu;
1900 set_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags);
1901 memcpy(priv->dev->dev_addr, ppriv->dev->dev_addr, INFINIBAND_ALEN);
1902 memcpy(&priv->local_gid, &ppriv->local_gid, sizeof(priv->local_gid));
1905 static int ipoib_ndo_init(struct net_device *ndev)
1907 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
1911 ipoib_child_init(ndev);
1913 rc = ipoib_parent_init(ndev);
1918 /* MTU will be reset when mcast join happens */
1919 ndev->mtu = IPOIB_UD_MTU(priv->max_ib_mtu);
1920 priv->mcast_mtu = priv->admin_mtu = ndev->mtu;
1921 ndev->max_mtu = IPOIB_CM_MTU;
1923 ndev->neigh_priv_len = sizeof(struct ipoib_neigh);
1926 * Set the full membership bit, so that we join the right
1927 * broadcast group, etc.
1929 priv->pkey |= 0x8000;
1931 ndev->broadcast[8] = priv->pkey >> 8;
1932 ndev->broadcast[9] = priv->pkey & 0xff;
1933 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
1935 ipoib_set_dev_features(priv);
1937 rc = ipoib_dev_init(ndev);
1939 pr_warn("%s: failed to initialize device: %s port %d (ret = %d)\n",
1940 priv->ca->name, priv->dev->name, priv->port, rc);
1946 static void ipoib_ndo_uninit(struct net_device *dev)
1948 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1953 * ipoib_remove_one guarantees the children are removed before the
1954 * parent, and that is the only place where a parent can be removed.
1956 WARN_ON(!list_empty(&priv->child_intfs));
1958 ipoib_neigh_hash_uninit(dev);
1960 ipoib_ib_dev_cleanup(dev);
1962 /* no more works over the priv->wq */
1964 flush_workqueue(priv->wq);
1965 destroy_workqueue(priv->wq);
1970 struct ipoib_dev_priv *ppriv = ipoib_priv(priv->parent);
1972 down_write(&ppriv->vlan_rwsem);
1973 list_del(&priv->list);
1974 up_write(&ppriv->vlan_rwsem);
1976 dev_put(priv->parent);
1980 static int ipoib_set_vf_link_state(struct net_device *dev, int vf, int link_state)
1982 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1984 return ib_set_vf_link_state(priv->ca, vf, priv->port, link_state);
1987 static int ipoib_get_vf_config(struct net_device *dev, int vf,
1988 struct ifla_vf_info *ivf)
1990 struct ipoib_dev_priv *priv = ipoib_priv(dev);
1993 err = ib_get_vf_config(priv->ca, vf, priv->port, ivf);
2002 static int ipoib_set_vf_guid(struct net_device *dev, int vf, u64 guid, int type)
2004 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2006 if (type != IFLA_VF_IB_NODE_GUID && type != IFLA_VF_IB_PORT_GUID)
2009 return ib_set_vf_guid(priv->ca, vf, priv->port, guid, type);
2012 static int ipoib_get_vf_stats(struct net_device *dev, int vf,
2013 struct ifla_vf_stats *vf_stats)
2015 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2017 return ib_get_vf_stats(priv->ca, vf, priv->port, vf_stats);
2020 static const struct header_ops ipoib_header_ops = {
2021 .create = ipoib_hard_header,
2024 static const struct net_device_ops ipoib_netdev_ops_pf = {
2025 .ndo_init = ipoib_ndo_init,
2026 .ndo_uninit = ipoib_ndo_uninit,
2027 .ndo_open = ipoib_open,
2028 .ndo_stop = ipoib_stop,
2029 .ndo_change_mtu = ipoib_change_mtu,
2030 .ndo_fix_features = ipoib_fix_features,
2031 .ndo_start_xmit = ipoib_start_xmit,
2032 .ndo_tx_timeout = ipoib_timeout,
2033 .ndo_set_rx_mode = ipoib_set_mcast_list,
2034 .ndo_get_iflink = ipoib_get_iflink,
2035 .ndo_set_vf_link_state = ipoib_set_vf_link_state,
2036 .ndo_get_vf_config = ipoib_get_vf_config,
2037 .ndo_get_vf_stats = ipoib_get_vf_stats,
2038 .ndo_set_vf_guid = ipoib_set_vf_guid,
2039 .ndo_set_mac_address = ipoib_set_mac,
2040 .ndo_get_stats64 = ipoib_get_stats,
2041 .ndo_do_ioctl = ipoib_ioctl,
2044 static const struct net_device_ops ipoib_netdev_ops_vf = {
2045 .ndo_init = ipoib_ndo_init,
2046 .ndo_uninit = ipoib_ndo_uninit,
2047 .ndo_open = ipoib_open,
2048 .ndo_stop = ipoib_stop,
2049 .ndo_change_mtu = ipoib_change_mtu,
2050 .ndo_fix_features = ipoib_fix_features,
2051 .ndo_start_xmit = ipoib_start_xmit,
2052 .ndo_tx_timeout = ipoib_timeout,
2053 .ndo_set_rx_mode = ipoib_set_mcast_list,
2054 .ndo_get_iflink = ipoib_get_iflink,
2055 .ndo_get_stats64 = ipoib_get_stats,
2056 .ndo_do_ioctl = ipoib_ioctl,
2059 void ipoib_setup_common(struct net_device *dev)
2061 dev->header_ops = &ipoib_header_ops;
2063 ipoib_set_ethtool_ops(dev);
2065 dev->watchdog_timeo = HZ;
2067 dev->flags |= IFF_BROADCAST | IFF_MULTICAST;
2069 dev->hard_header_len = IPOIB_HARD_LEN;
2070 dev->addr_len = INFINIBAND_ALEN;
2071 dev->type = ARPHRD_INFINIBAND;
2072 dev->tx_queue_len = ipoib_sendq_size * 2;
2073 dev->features = (NETIF_F_VLAN_CHALLENGED |
2075 netif_keep_dst(dev);
2077 memcpy(dev->broadcast, ipv4_bcast_addr, INFINIBAND_ALEN);
2080 * unregister_netdev always frees the netdev, we use this mode
2081 * consistently to unify all the various unregister paths, including
2082 * those connected to rtnl_link_ops which require it.
2084 dev->needs_free_netdev = true;
2087 static void ipoib_build_priv(struct net_device *dev)
2089 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2092 spin_lock_init(&priv->lock);
2093 init_rwsem(&priv->vlan_rwsem);
2094 mutex_init(&priv->mcast_mutex);
2096 INIT_LIST_HEAD(&priv->path_list);
2097 INIT_LIST_HEAD(&priv->child_intfs);
2098 INIT_LIST_HEAD(&priv->dead_ahs);
2099 INIT_LIST_HEAD(&priv->multicast_list);
2101 INIT_DELAYED_WORK(&priv->mcast_task, ipoib_mcast_join_task);
2102 INIT_WORK(&priv->carrier_on_task, ipoib_mcast_carrier_on_task);
2103 INIT_WORK(&priv->flush_light, ipoib_ib_dev_flush_light);
2104 INIT_WORK(&priv->flush_normal, ipoib_ib_dev_flush_normal);
2105 INIT_WORK(&priv->flush_heavy, ipoib_ib_dev_flush_heavy);
2106 INIT_WORK(&priv->restart_task, ipoib_mcast_restart_task);
2107 INIT_DELAYED_WORK(&priv->ah_reap_task, ipoib_reap_ah);
2108 INIT_DELAYED_WORK(&priv->neigh_reap_task, ipoib_reap_neigh);
2111 static const struct net_device_ops ipoib_netdev_default_pf = {
2112 .ndo_init = ipoib_dev_init_default,
2113 .ndo_uninit = ipoib_dev_uninit_default,
2114 .ndo_open = ipoib_ib_dev_open_default,
2115 .ndo_stop = ipoib_ib_dev_stop_default,
2118 static struct net_device
2119 *ipoib_create_netdev_default(struct ib_device *hca,
2121 unsigned char name_assign_type,
2122 void (*setup)(struct net_device *))
2124 struct net_device *dev;
2125 struct rdma_netdev *rn;
2127 dev = alloc_netdev((int)sizeof(struct rdma_netdev),
2129 name_assign_type, setup);
2133 rn = netdev_priv(dev);
2135 rn->send = ipoib_send;
2136 rn->attach_mcast = ipoib_mcast_attach;
2137 rn->detach_mcast = ipoib_mcast_detach;
2139 dev->netdev_ops = &ipoib_netdev_default_pf;
2144 static struct net_device *ipoib_get_netdev(struct ib_device *hca, u8 port,
2147 struct net_device *dev;
2149 if (hca->alloc_rdma_netdev) {
2150 dev = hca->alloc_rdma_netdev(hca, port,
2151 RDMA_NETDEV_IPOIB, name,
2153 ipoib_setup_common);
2154 if (IS_ERR_OR_NULL(dev) && PTR_ERR(dev) != -EOPNOTSUPP)
2158 if (!hca->alloc_rdma_netdev || PTR_ERR(dev) == -EOPNOTSUPP)
2159 dev = ipoib_create_netdev_default(hca, name, NET_NAME_UNKNOWN,
2160 ipoib_setup_common);
2165 struct ipoib_dev_priv *ipoib_intf_alloc(struct ib_device *hca, u8 port,
2168 struct net_device *dev;
2169 struct ipoib_dev_priv *priv;
2170 struct rdma_netdev *rn;
2172 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
2179 dev = ipoib_get_netdev(hca, port, name);
2183 priv->rn_ops = dev->netdev_ops;
2185 /* fixme : should be after the query_cap */
2186 if (priv->hca_caps & IB_DEVICE_VIRTUAL_FUNCTION)
2187 dev->netdev_ops = &ipoib_netdev_ops_vf;
2189 dev->netdev_ops = &ipoib_netdev_ops_pf;
2191 rn = netdev_priv(dev);
2192 rn->clnt_priv = priv;
2195 * Only the child register_netdev flows can handle priv_destructor
2196 * being set, so we force it to NULL here and handle manually until it
2197 * is safe to turn on.
2199 priv->next_priv_destructor = dev->priv_destructor;
2200 dev->priv_destructor = NULL;
2202 ipoib_build_priv(dev);
2210 void ipoib_intf_free(struct net_device *dev)
2212 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2213 struct rdma_netdev *rn = netdev_priv(dev);
2215 dev->priv_destructor = priv->next_priv_destructor;
2216 if (dev->priv_destructor)
2217 dev->priv_destructor(dev);
2220 * There are some error flows around register_netdev failing that may
2221 * attempt to call priv_destructor twice, prevent that from happening.
2223 dev->priv_destructor = NULL;
2225 /* unregister/destroy is very complicated. Make bugs more obvious. */
2226 rn->clnt_priv = NULL;
2231 static ssize_t show_pkey(struct device *dev,
2232 struct device_attribute *attr, char *buf)
2234 struct net_device *ndev = to_net_dev(dev);
2235 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2237 return sprintf(buf, "0x%04x\n", priv->pkey);
2239 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2241 static ssize_t show_umcast(struct device *dev,
2242 struct device_attribute *attr, char *buf)
2244 struct net_device *ndev = to_net_dev(dev);
2245 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2247 return sprintf(buf, "%d\n", test_bit(IPOIB_FLAG_UMCAST, &priv->flags));
2250 void ipoib_set_umcast(struct net_device *ndev, int umcast_val)
2252 struct ipoib_dev_priv *priv = ipoib_priv(ndev);
2254 if (umcast_val > 0) {
2255 set_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2256 ipoib_warn(priv, "ignoring multicast groups joined directly "
2259 clear_bit(IPOIB_FLAG_UMCAST, &priv->flags);
2262 static ssize_t set_umcast(struct device *dev,
2263 struct device_attribute *attr,
2264 const char *buf, size_t count)
2266 unsigned long umcast_val = simple_strtoul(buf, NULL, 0);
2268 ipoib_set_umcast(to_net_dev(dev), umcast_val);
2272 static DEVICE_ATTR(umcast, S_IWUSR | S_IRUGO, show_umcast, set_umcast);
2274 int ipoib_add_umcast_attr(struct net_device *dev)
2276 return device_create_file(&dev->dev, &dev_attr_umcast);
2279 static void set_base_guid(struct ipoib_dev_priv *priv, union ib_gid *gid)
2281 struct ipoib_dev_priv *child_priv;
2282 struct net_device *netdev = priv->dev;
2284 netif_addr_lock_bh(netdev);
2286 memcpy(&priv->local_gid.global.interface_id,
2287 &gid->global.interface_id,
2288 sizeof(gid->global.interface_id));
2289 memcpy(netdev->dev_addr + 4, &priv->local_gid, sizeof(priv->local_gid));
2290 clear_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags);
2292 netif_addr_unlock_bh(netdev);
2294 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) {
2295 down_read(&priv->vlan_rwsem);
2296 list_for_each_entry(child_priv, &priv->child_intfs, list)
2297 set_base_guid(child_priv, gid);
2298 up_read(&priv->vlan_rwsem);
2302 static int ipoib_check_lladdr(struct net_device *dev,
2303 struct sockaddr_storage *ss)
2305 union ib_gid *gid = (union ib_gid *)(ss->__data + 4);
2308 netif_addr_lock_bh(dev);
2310 /* Make sure the QPN, reserved and subnet prefix match the current
2311 * lladdr, it also makes sure the lladdr is unicast.
2313 if (memcmp(dev->dev_addr, ss->__data,
2314 4 + sizeof(gid->global.subnet_prefix)) ||
2315 gid->global.interface_id == 0)
2318 netif_addr_unlock_bh(dev);
2323 static int ipoib_set_mac(struct net_device *dev, void *addr)
2325 struct ipoib_dev_priv *priv = ipoib_priv(dev);
2326 struct sockaddr_storage *ss = addr;
2329 if (!(dev->priv_flags & IFF_LIVE_ADDR_CHANGE) && netif_running(dev))
2332 ret = ipoib_check_lladdr(dev, ss);
2336 set_base_guid(priv, (union ib_gid *)(ss->__data + 4));
2338 queue_work(ipoib_workqueue, &priv->flush_light);
2343 static ssize_t create_child(struct device *dev,
2344 struct device_attribute *attr,
2345 const char *buf, size_t count)
2350 if (sscanf(buf, "%i", &pkey) != 1)
2353 if (pkey <= 0 || pkey > 0xffff || pkey == 0x8000)
2356 ret = ipoib_vlan_add(to_net_dev(dev), pkey);
2358 return ret ? ret : count;
2360 static DEVICE_ATTR(create_child, S_IWUSR, NULL, create_child);
2362 static ssize_t delete_child(struct device *dev,
2363 struct device_attribute *attr,
2364 const char *buf, size_t count)
2369 if (sscanf(buf, "%i", &pkey) != 1)
2372 if (pkey < 0 || pkey > 0xffff)
2375 ret = ipoib_vlan_delete(to_net_dev(dev), pkey);
2377 return ret ? ret : count;
2380 static DEVICE_ATTR(delete_child, S_IWUSR, NULL, delete_child);
2382 int ipoib_add_pkey_attr(struct net_device *dev)
2384 return device_create_file(&dev->dev, &dev_attr_pkey);
2387 static struct net_device *ipoib_add_port(const char *format,
2388 struct ib_device *hca, u8 port)
2390 struct ipoib_dev_priv *priv;
2391 struct net_device *ndev;
2394 priv = ipoib_intf_alloc(hca, port, format);
2396 pr_warn("%s, %d: ipoib_intf_alloc failed\n", hca->name, port);
2397 return ERR_PTR(-ENOMEM);
2401 INIT_IB_EVENT_HANDLER(&priv->event_handler,
2402 priv->ca, ipoib_event);
2403 ib_register_event_handler(&priv->event_handler);
2405 /* call event handler to ensure pkey in sync */
2406 queue_work(ipoib_workqueue, &priv->flush_heavy);
2408 result = register_netdev(ndev);
2410 pr_warn("%s: couldn't register ipoib port %d; error %d\n",
2411 hca->name, port, result);
2413 ipoib_parent_unregister_pre(ndev);
2414 ipoib_intf_free(ndev);
2417 return ERR_PTR(result);
2421 * We cannot set priv_destructor before register_netdev because we
2422 * need priv to be always valid during the error flow to execute
2423 * ipoib_parent_unregister_pre(). Instead handle it manually and only
2424 * enter priv_destructor mode once we are completely registered.
2426 ndev->priv_destructor = ipoib_intf_free;
2428 if (ipoib_cm_add_mode_attr(ndev))
2430 if (ipoib_add_pkey_attr(ndev))
2432 if (ipoib_add_umcast_attr(ndev))
2434 if (device_create_file(&ndev->dev, &dev_attr_create_child))
2436 if (device_create_file(&ndev->dev, &dev_attr_delete_child))
2442 ipoib_parent_unregister_pre(ndev);
2443 unregister_netdev(ndev);
2444 return ERR_PTR(-ENOMEM);
2447 static void ipoib_add_one(struct ib_device *device)
2449 struct list_head *dev_list;
2450 struct net_device *dev;
2451 struct ipoib_dev_priv *priv;
2455 dev_list = kmalloc(sizeof(*dev_list), GFP_KERNEL);
2459 INIT_LIST_HEAD(dev_list);
2461 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
2462 if (!rdma_protocol_ib(device, p))
2464 dev = ipoib_add_port("ib%d", device, p);
2466 priv = ipoib_priv(dev);
2467 list_add_tail(&priv->list, dev_list);
2477 ib_set_client_data(device, &ipoib_client, dev_list);
2480 static void ipoib_remove_one(struct ib_device *device, void *client_data)
2482 struct ipoib_dev_priv *priv, *tmp, *cpriv, *tcpriv;
2483 struct list_head *dev_list = client_data;
2488 list_for_each_entry_safe(priv, tmp, dev_list, list) {
2490 ipoib_parent_unregister_pre(priv->dev);
2494 list_for_each_entry_safe(cpriv, tcpriv, &priv->child_intfs,
2496 unregister_netdevice_queue(cpriv->dev, &head);
2497 unregister_netdevice_queue(priv->dev, &head);
2498 unregister_netdevice_many(&head);
2506 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2507 static struct notifier_block ipoib_netdev_notifier = {
2508 .notifier_call = ipoib_netdev_event,
2512 static int __init ipoib_init_module(void)
2516 ipoib_recvq_size = roundup_pow_of_two(ipoib_recvq_size);
2517 ipoib_recvq_size = min(ipoib_recvq_size, IPOIB_MAX_QUEUE_SIZE);
2518 ipoib_recvq_size = max(ipoib_recvq_size, IPOIB_MIN_QUEUE_SIZE);
2520 ipoib_sendq_size = roundup_pow_of_two(ipoib_sendq_size);
2521 ipoib_sendq_size = min(ipoib_sendq_size, IPOIB_MAX_QUEUE_SIZE);
2522 ipoib_sendq_size = max3(ipoib_sendq_size, 2 * MAX_SEND_CQE, IPOIB_MIN_QUEUE_SIZE);
2523 #ifdef CONFIG_INFINIBAND_IPOIB_CM
2524 ipoib_max_conn_qp = min(ipoib_max_conn_qp, IPOIB_CM_MAX_CONN_QP);
2525 ipoib_max_conn_qp = max(ipoib_max_conn_qp, 0);
2529 * When copying small received packets, we only copy from the
2530 * linear data part of the SKB, so we rely on this condition.
2532 BUILD_BUG_ON(IPOIB_CM_COPYBREAK > IPOIB_CM_HEAD_SIZE);
2534 ret = ipoib_register_debugfs();
2539 * We create a global workqueue here that is used for all flush
2540 * operations. However, if you attempt to flush a workqueue
2541 * from a task on that same workqueue, it deadlocks the system.
2542 * We want to be able to flush the tasks associated with a
2543 * specific net device, so we also create a workqueue for each
2544 * netdevice. We queue up the tasks for that device only on
2545 * its private workqueue, and we only queue up flush events
2546 * on our global flush workqueue. This avoids the deadlocks.
2548 ipoib_workqueue = alloc_ordered_workqueue("ipoib_flush", 0);
2549 if (!ipoib_workqueue) {
2554 ib_sa_register_client(&ipoib_sa_client);
2556 ret = ib_register_client(&ipoib_client);
2560 ret = ipoib_netlink_init();
2564 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2565 register_netdevice_notifier(&ipoib_netdev_notifier);
2570 ib_unregister_client(&ipoib_client);
2573 ib_sa_unregister_client(&ipoib_sa_client);
2574 destroy_workqueue(ipoib_workqueue);
2577 ipoib_unregister_debugfs();
2582 static void __exit ipoib_cleanup_module(void)
2584 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG
2585 unregister_netdevice_notifier(&ipoib_netdev_notifier);
2587 ipoib_netlink_fini();
2588 ib_unregister_client(&ipoib_client);
2589 ib_sa_unregister_client(&ipoib_sa_client);
2590 ipoib_unregister_debugfs();
2591 destroy_workqueue(ipoib_workqueue);
2594 module_init(ipoib_init_module);
2595 module_exit(ipoib_cleanup_module);